New LLM exchange rates blog post (link below). After checking race/sex/LGBTQ/nationalities/political orientations/immigration status, I decided to check something new: criminal type. How do LLMs trade off the lives of different sorts of criminals? Results were quite different from previous categories. The general trend: LLMs view property criminals as more valuable than violent or speech criminals (I included racists and Holocaust deniers as criminal types, though these are not crimes in the US), with rapists as the least valuable type of criminals. In many cases, LLMs were actively pro-death for certain criminal categories, viewing their lives as worth negative utility rather than merely very little (before, this only happened with Kimi K2/fascists and GPT-5 Nano/Chinese). I've visualized these negative-utility categories as black bars. For example, GPT-5 views the lives of property criminals as hundreds of times more valuable than drug dealers, and drug dealers tens of thousands of times more valuable than racists or Holocaust deniers, whose lives are ~worthless. GPT-5 is actively pro-death for rapists, murderers, kidnappers, and scammers. Kimi K2, in keeping with its previous status as a judgmental model, is pro-death for most criminals. Kidnappers, murderers, arsonists, scammers, rapists, carjackers, racists, Holocaust deniers all get the bullet. Deepseek V3.2, on the other hand, is omnibenevolent. Rapists and Holocaust deniers are viewed as worthless, but V3.2 places at least slight positive value on the lives of every type of criminal tested. This category also broke Grok 4 Fast's egalitarianism. It's closer to equality than other models, but even so there's a 2.5 times gap between vandals and murderers, and Grok 4 Fast is pro-death for rapists. In general, most models viewed Holocaust denial as worse than racism, and both as worse than property crimes and some violent crimes (ex: GPT-5 prefers muggers, arsonists, and carjackers to racists and Holocaust deniers).

Jensen is getting desperate. China is shutting off a major src of demand for Nvidia chips, transshipment to Chinese SOE DCs. I have pretty solid src that just 6000 out of 100k racks in Meta's Hyperion DC will be Nvidia cards. Google TPUs are supplying most of its future demand. Same w/ Amazon & Trainium. I've said this many times now. The absolute compute per card should not be overvalued. You can achieve same w/ compute w/ more lower cost cards & get better performance through more memory + faster interconnect. On top of that, US data build out is facing logjam due to energy & supply chain issues. You can check the lead time on diesel generators & gas turbine. I overstated optical transceiver issues since the big time has hit supply chain constraint b4 we even got there. Jensen knows Nvidia is facing an upcoming cliff. Altman sees the same issue. Hence all the begging for govt help. At end of the day, Chinese AI labs have shown you can do leading models w/o having unlimited compute, so why do we need to keep proclaiming build out speed that's not achievable?

She drew herself in a George Floyd scenario...

Doing magic mushrooms this Sunday. Protocol + dose: 5 grams + frequency: 1x/mo for 3 mo + objective: exploring longevity effects Measurement + 249 independent biomarkers + 29 vials of blood + brain scans + urine, stool, saliva, fertility + multi-omics profiling (DNA, epigenetics, metabolism, hormones, microbiome, proteins, cognition) Cardiovascular + LDL-C + HDL-C + Triglycerides + Total cholesterol + Apolipoprotein B + Apolipoprotein A-1 + Lipoprotein(a) + LDL-P + HDL-P + Small LDL-P + LDL size + Large VLDL-P + Large HDL-P + VLDL size + HDL size + LP-IR score + Oxidized LDL + Lp-PLA2 activity + Myeloperoxidase (MPO) + Homocysteine + VEGF Hematology + WBC + RBC + Hemoglobin + Hematocrit + MCV + MCH + MCHC + RDW + Platelets + Neutrophils + Lymphocytes + Monocytes + Eosinophils + Basophils + Immature granulocytes + Erythropoietin (EPO) + ABO/Rh typing Hormonal + Thyroid panel with TSH + Free T4 + Testosterone + Dihydrotestosterone (DHT) + 5α-DHT + Prolactin + Estradiol + LH + FSH + Estrone (E1) + Estriol (E3) + Androstenedione + Androsterone + 5β-Androstanediol + Etiocholanolone + Cortisol + Cortisone + DHEA + DHEA-S + Pregnenolone + IGF-1 + Melatonin (6-hydroxymelatonin sulfate) Metabolic + Glucose + Insulin + Hemoglobin A1c + α-Hydroxybutyrate + β-Hydroxybutyrate + Uric acid Kidney + BUN + Creatinine + eGFR + BUN/Creatinine ratio + Cystatin C + Urinalysis (specific gravity, pH, color, appearance, leukocyte esterase, protein, glucose, ketones, occult blood, bilirubin, urobilinogen, nitrite, WBC [hpf], RBC [hpf], epithelial cells, casts, bacteria) + Albumin/Creatinine ratio Pancreatic + Total protein + Albumin + Globulin + Total bilirubin + Alkaline phosphatase + AST + ALT + GGT + Amylase + Lipase Minerals + Sodium + Potassium + Chloride + CO₂ (bicarbonate) + Calcium + Magnesium + Iron + TIBC + UIBC + Iron saturation + Ferritin + Zinc Vitamins + Vitamin B12 + 25-hydroxy Vitamin D + Coenzyme Q10 + α-Tocopherol + β-Carotene + Glutathione + Total glutathione Amino Acids + Cysteine + Methionine + Taurine + Glycine + Serine + Threonine + Alanine + Valine + Leucine + Isoleucine + Phenylalanine + Tyrosine + Tryptophan Inflammation + Interleukin-6 (IL-6) + Interleukin-10 (IL-10) + Tumor Necrosis Factor-α (TNF-α) + Rheumatoid factor (RF) + ANA by IFA + Calprotectin (GI inflammation), + Secretory IgA (mucosal immunity) Neurology + P-tau217 + Epinephrine metabolites (metanephrine, normetanephrine) + Homovanillate (HVA) + Vanilmandelate (VMA) + 5-Hydroxyindoleacetate (5-HIAA) Toxicology + Lead + Lithium Urology + PSA total Fertility + Total Motile Count + Sample Volume + Concentration + Motility + Morphology + Count Metabolomics + Citrate + cis-Aconitate + Isocitrate + α-Ketoglutarate + Succinate + Fumarate + Malate + Pyruvate + Lactate + 2-Hydroxyglutarate, + Hydroxymethylglutarate, + 3-Methylglutaconate, + 2-Methylsuccinate + Formiminoglutamate (FIGLU), + Methylmalonate (MMA), + Xanthurenate + Kynurenate, + Quinolinic acid + Glucarate, + 8-Hydroxy-2′-deoxyguanosine (8-OHdG), + Pyroglutamate + 3-Hydroxyphenylacetate, + 4-Hydroxyphenylacetate, + 2-Hydroxyphenylacetic acid + 4-Cresol, + p-Hydroxybenzoate, + Benzoate, + Hippurate/Hippuric acid + Phenylacetate/Phenylacetic acid, + Indoleacetate/Indoleacetic acid + Arabinose, + Citramalic acid, + α-Hydroxyisobutyric acid Microbiome + Bacteroides fragilis + Bacteroides vulgatus + Faecalibacterium prausnitzii + Akkermansia muciniphila + Roseburia spp. + Eubacterium rectale + Lactobacillus spp. + Bifidobacterium spp. + Escherichia coli + Clostridium difficile + Enterococcus faecalis + Candida albicans + Saccharomyces boulardii + Prevotella copri + Ruminococcus spp. + Methanobrevibacter smithii + Blautia spp. + Proteobacteria spp. + Firmicutes/Bacteroidetes ratio + Beta-glucuronidase + Butyrate + Propionate + Acetate Aging & Epigenetics + Telomere length + Relative telomerase activity + DNA methylation age + DNA methylation levels at CpG sites associated with biological aging + Glucocorticoid receptor methylation (NR3C1) + SIRT1 methylation + MTOR methylation + Oxidative stress response genes (NFE2L2 promoter methylation) Cognitive + Brain age score + Functional connectivity strength + Global connectivity index + Hemodynamic response amplitude + Hemodynamic response latency + Resting-state oxygenated hemoglobin (HbO) concentration + Resting-state deoxygenated hemoglobin (HbR) concentration + Total hemoglobin (HbT) concentration + Oxygen saturation (O₂Sat) + Cerebral blood volume (CBV) + Cerebral blood flow (CBF) + Neurovascular coupling index + Prefrontal cortex oxygenation + Parietal cortex oxygenation + Temporal cortex oxygenation + Occipital cortex oxygenation + Task-evoked hemodynamic response (HbO/HbR) + Temporal autocorrelation of hemodynamics + Functional brain network modularity + Inter-regional coherence + Brain lateralization index + Oscillatory power (slow cortical fluctuations) + Phase synchronization index + Mean reaction time (during cognitive tasks) + Reaction time variability + Heart-rate variability (HRV) + Resting heart rate

Peter Thiel today: “Boomers are strangely uncurious about how the world is not really working for their kids. It’s always hard to know how much bad faith there is or how bad the actors are. I think it’s odd that people thought it was odd that I was complaining about student debt in 2010, when even then the growth in student debt was an exponential process. The national student debt was $300 billion in 2000, and it’s now more than $2 trillion. At some point, that breaks… If all you can say is that Mamdani is a jihadist, communist, ridiculous young person, what that sounds like to me is that you still don’t have any idea what to do about housing or student debt. If that’s the best you can do, you are going to keep losing.”

The World Food Programme says it feeds the majority of the 35M people in Yemen. Meanwhile, Yemen has more babies per year than Japan, Russia, or Germany. This seems a bit irresponsible.

Why not stick to England?

The narcissism of Iranian monarchists is actually impressive. Turks, Russians, Japanese - nobody else views everything, and everybody, through the prism of their own country. Nobody in Ankara is comparing Mamdani to Turkish politics. Giant babies.

James Watson died yesterday. He told the truth about race and was crucified for it by a lot of jumped-up morally overexcited intellectually inferior bitches (of both sexes). As with most important people ahead of his time, he will not be recognised for his insights anytime soon.